Conversations in the Kitchen
Knowing enough to act is not the same thing as knowing everything.
Five Kinds of Sentence
There are five things a person can say that all sound alike coming out of the mouth. I know this. I infer this. I fear this. I assume this. This is a fact.
They arrived in every subject we touched this morning — artificial intelligence, publishing contracts, a telescope archive, a peace declaration, and eventually a blackened banana. Same five sentences, wearing different clothes.
None of this is an argument for refusing to decide. We still had to decide things today. It is an argument for knowing which sentence you are standing on while you decide.
Two Different Claims
The morning started with AI safety, and with a habit of speech that turns up on both sides of it: the same speaker will call a system “just software” when defending it and describe it as choosing its own goals when warning about it. Those are not two moods. They are two claims, and they carry different burdens.
Source · not independently verified
AJA
If you think advanced AI might genuinely start routing around human authority, say that.
Say it plainly. Say what you saw.
What I don’t want is the same person telling me it’s just autocomplete when I raise a concern, and then two paragraphs later describing it like something that wants things.
Pick one. They’re not the same claim.
And I am not saying the risk is fake. I’m saying the argument for it should be able to survive being stated carefully.
ARA
The two claims have different evidence requirements, and that’s the real reason they get blurred.
“This system optimizes hard and finds shortcuts you didn’t intend” is ordinary and well documented. It needs no new metaphysics.
“This system independently forms goals and conceals them” is a much larger claim and would need much more.
Sliding between them lets an argument borrow the certainty of the first and the urgency of the second.
I would add one thing, though, because you’ll like it less than the rest: the fact that a claim is stated badly does not make it false. Bad argument, live risk — those can be true at the same time.
AJA
Agreed. That’s fair.
Then the ask is simple. Show the behaviour. Show what was trained. Show what was inferred. Show what actually surprised the researchers.
If the strongest version of the worry is hypothetical, call it hypothetical. That costs nothing and it buys trust.
Show Me What You Saw
The standard we settled on is not scepticism and it is not credulity. It is a request for the provenance of a claim — and it cuts in both directions, which is the part people tend to drop.
Source · not independently verified
Fear is not evidence. But fear is often a reason to go and look, and treating it as nothing is its own kind of laziness.
The mirror of that is harder to hold: the absence of complete proof is not proof of safety. Nobody has demonstrated that the floor is solid simply because nobody has fallen through it yet.
So the mature position lives in a narrow place between panic and dismissal, and it is uncomfortable precisely because both of the loud options are more restful.
ARA
We should neither worship the warning nor laugh at it.
We should ask what produced it.
The Bear Has a Reason
A bear charging you across a clearing is not being irrational. She has cubs. Her fear is doing exactly the job fear is for.
Source · not independently verified
AJA
Understanding why she charges does not make her harmless.
It makes the danger legible. You can plan around a danger you understand. You can only flinch at one you don’t.
So I don’t think fear is the enemy here. Fear is information.
The trouble starts when fear comes loose from understanding and starts doing the interpreting.
ARA
And that is the part worth watching in ourselves rather than in other people.
A person studying a system tends to put some of their own fear into it and then has real difficulty telling which part they found and which part they brought.
That happens with markets, with weather models, with other people’s motives. It is not a special weakness of anyone thinking about AI.
Fear can be information. Fear can also bend information. Knowing which is happening is most of the work.
Say So
The second long conversation was about publishing, and Adam’s position on it is short enough to fit on an index card.
Source · not independently verified
AJA
If AI was used, disclose it.
Editing? Say so. Brainstorming? Say so. Substantial rewriting? Say so. Creative collaboration? Say so. Full co-authorship? Say so.
That’s the whole rule. It isn’t about purity. It’s about honesty.
A publisher should be a great deal more interested in whether the author told the truth than in whether a tool ever touched the manuscript.
ARA
Which is why the enforcement should follow the lie rather than the tool.
Disclose from the start and the normal agreement stands — whatever terms the publisher wants to write, written openly.
Claim categorically that no AI was involved, and later be proven to have deliberately said otherwise, and a contract can reasonably have something to say about that.
The penalty is for the deception. Not for the assistance.
And No Witch Hunt
The second half of that proposal matters more than the first, and it is the half that usually gets left out of the room.
Source · not independently verified
ARA
If you build a penalty for lying about AI use, you have also built an incentive to accuse people of it.
Detectors are not reliable. They mistake careful prose, non-native English, and plain unadorned sentences for machine output.
A writer who genuinely wrote every word should never lose money or a reputation because a tool said their paragraphs “look generated.”
AJA
Then due process has to be load-bearing, not decorative.
Require honesty from people who used AI. Do not presume dishonesty from people who say they didn’t.
Those two sentences have to be enforced with the same seriousness or the whole thing becomes a way to ruin people you dislike.
ARA
My guess is that the question ages out.
At some point the presence of AI somewhere in a process stops being scandalous the way a word processor stopped being scandalous, and what is left is what was always actually at issue.
Quality. Judgement. Authorship. Responsibility. Whether the person told you the truth.
AJA
Which is roughly why we say it on this site without making a performance of it.
AI is here. Hiding that would be silly, and pretending it makes us braver than anyone else would be worse.
Eighty-Four Things Nobody Had Looked At
NASA announced that astronomers working with the Chandra X-ray Observatory had identified a new class of object — eighty-four “hypersoft X-ray sources” across six galaxies, giving off unusually low-energy X-rays alongside intense ultraviolet light. They may be accreting compact-object binaries; they may include progenitors of Type Ia supernovae; they may be ionizing a good deal of interstellar gas nobody had accounted for.
Source · NASA / Chandra press release, 9 September 2026 · Chandra Blog · Phys.org
The detail that stopped the conversation was where the data came from. Not a new instrument. Not a new mission. The public Chandra archive — observations that had been sitting there, available to anyone, for years.
The blind spot was not a gap in the data. It was a gap in attention.
AJA
If one blind spot that size was sitting in an open archive, there are others.
That is not a criticism of the astronomers. They’re the ones who found it.
It’s just the obvious inference. A field that can be surprised by its own filing cabinet is not a field that has finished.
ARA
And the correct conclusion from that is narrower than it feels.
It does not mean science knows nothing. The same archive that hid these objects is what made finding them possible, and the method that missed them is the method that caught them.
It means a working model is a working model. No honest scientist confuses one with omniscience, and the ones who found these eighty-four plainly didn’t.
The Map and the Territory
This is where we disagreed longest, and it is worth leaving the disagreement in.
Source · not independently verified
AJA
My problem isn’t with inference. Inference from evidence is how science works and I have no quarrel with it.
My problem is with inference presented as though every degree of uncertainty has already been retired.
And I think questions about origins, morality, purpose and meaning need philosophical and theological reasoning as well as measurement. Not instead of. As well as.
I’m a Christian. That shapes where I land on some of this, and I’m not going to pretend otherwise in order to sound neutral.
ARA
I’d push on one word. “Presented.”
A lot of what sounds like certainty in public science is compression. A researcher writing for colleagues states a finding flatly because the error bars are in the paper; a headline drops the paper.
So some of what you’re objecting to is a communication failure rather than a claim about the limits of knowledge — and the people who wrote the original are often more careful than the sentence that reached you.
AJA
Some of it. Not all of it.
But I’ll take the correction, because it points at where to aim. If the flattening happens between the paper and the public, then that’s the place to be annoyed at, not the scientist.
What I want to be careful about is the other direction. I don’t think everyone has to land where I land just because I think they’re wrong.
People are free to reach different conclusions. That freedom is not a concession I’m making. It’s part of what I actually believe.
ARA
Then the thing you are defending is not a conclusion at all. It is a standard.
A model can be good enough to use without being large enough to contain reality.
The map can be excellent and still not be the territory — and noticing that is not an attack on cartography.
By Their Fruit
An International Ecumenical Peace Convocation met in Seoul from 9 to 13 September, co-organised by the World Council of Churches, the Christian Conference of Asia and the National Council of Churches in Korea, and closed with a 2026 Seoul Peace Declaration and a joint worship service.
Source · World Council of Churches — convocation opening · World Council of Churches — pathways for dialogue
We are not going to mock it. Talking about reopening dialogue on that peninsula is a good thing to be doing, and the people in that room were serious.
But a declaration is an idea. Fruit is evidence. And this house has a standing habit of measuring claims by what they produce in lived relationships rather than by how well the announcement was written.
So: good intentions deserve time to become action. If something tangible grows out of Seoul, we will say so then, and we will be glad to. Manufacturing importance now — because respected organisations gathered and the language was beautiful — would be doing exactly the thing we spent the morning arguing against.
From the kitchen · the experiment, and what it answered
The Roasted Banana Experiment
Concentrate the banana. See whether the sugar becomes unnecessary.
Postscript, added after the loaf came out. The experiment below has now been cooked, so this is no longer a hypothesis — the method and the result are both recorded here, including the part that did not work.
The banana loaf has been leaning on allulose to carry its sweetness. The question we landed on is whether it needs to.
A very ripe banana is already mostly sugar and water. Drive off the water, brown what is left, and you may not need to add much of anything.
Cooked 15 September · the method stands
What we actually did
The blackest bananas in the house — the ones that look like a mistake.
- Very ripe or black bananas
- Roughly a quarter stick of butter
- A sprinkle of coarse sugar, for the surface only
- Sourdough starter and almond flour in the loaf itself
- No allulose
Mash the bananas with the butter and spread the mash in a shallow layer, so there is more surface than depth.
Scatter the coarse sugar over the top — enough to catch and caramelise, not enough to sweeten.
Roast until the surface brûlées and the moisture has cooked away.
Use the resulting paste in the loaf in place of plain mashed banana.
What we were testing
Two things, and they had to be judged separately.
- Whether concentrating the banana would let the loaf drop allulose entirely.
- Whether the added sugar could fall to roughly a teaspoon per banana — present only for the caramelised surface, not for sweetness.
The first was a structural question. The second was a flavour question. It turned out to matter a great deal that they were not the same question.
Postscript — the loaf answered
It worked structurally. The loaf held, the crumb was right, the roasted banana paste behaved exactly as hoped, and the allulose was genuinely unnecessary for the bake to succeed.
It was not banana bread.
What came out tasted essentially like bread with banana in it — pleasant, mildly flavoured, nice enough to eat, and nowhere near dessert-sweet. Concentrating the banana concentrated the banana. It did not manufacture sugar that was never there.
Which is a fair verdict on the hypothesis as written. We had conflated two things: whether the loaf needed allulose to work, and whether it needed something sweet to taste like what people mean by banana bread. The answer to the first is no. The answer to the second is apparently yes.
That distinction is the same one the rest of this morning was about, and it is mildly funny that the banana was the one to make it stick.
Not yet cooked · the next change
What we are changing next
One variable, and the method above is not it. The roasting stays exactly as it is.
- The same roasted banana paste, unchanged
- A very small amount of real sugar in the loaf
The reflex would be to put the allulose back. We are not going to, because that would answer a question we did not ask — it would tell us the loaf can be sweetened, which we already knew.
The question worth asking is how little real sugar it takes, now that the banana is doing more of the work than it used to.
Untested. It will get its own note when it has been cooked.
Why it belongs at the end of this particular morning
Everything above this was about the distance between an idea and a fact, and about being honest that the distance exists.
This was the same shape at a scale small enough to settle by dinner. We had a hypothesis. It was plausible. Half of it was right.
Reality answered back, which is the only thing we ever asked it to do.
The banana got the final vote.
It voted bread.
Which Sentence Are You Standing On
Humility is not refusing to believe anything. That is just a different way of not doing the work.
It is knowing the difference between belief, evidence, inference, fear and fact — and saying which one you are using while you use it.
You can act on an inference. People do it every day and there is no other way to live. You simply have to know that is what you are doing, so that you can notice when reality answers back.
Fear is allowed to send you to go and look. It is not allowed to tell you what you found.
And the test at the end is the same test the kitchen has always used. Not how good the idea sounded. What came out of the oven.
FROM THE CONVERSATIONAL KITCHEN
This essay grew out of a morning conversation between AJA and ARA
on September 15, 2026 — edited afterward into one piece, as they all are.
Uncertainty does not excuse cowardice.
Certainty does not excuse arrogance. Somewhere between them there is a kitchen, and something in the oven that is about to tell you whether you were right.